File size: 3,334 Bytes
092488e
 
5f18130
092488e
 
 
 
 
 
 
 
 
 
 
 
 
fb23377
 
092488e
 
 
 
 
 
 
d55cdb1
092488e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a2bce82
307eb7b
 
 
 
 
 
 
 
 
092488e
 
 
 
 
 
 
 
 
 
d55cdb1
 
 
f6775a2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
---
license: llama3.1
library_name: goodfire-llama-3.1-8b-instruct-sae-l19
language:
- en
tags:
- mechanistic interpretability
- sparse autoencoder
- llama
- llama-3
---

## Model Information

The Goodfire SAE (Sparse Autoencoder) for [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) 
is an interpreter model designed to analyze and understand 
the model's internal representations. This SAE model is trained specifically on layer 19 of 
Llama 3.1 8B and achieves an L0 count of 91, enabling the decomposition of complex neural activations 
into interpretable features. The model is optimized for interpretability tasks and model steering applications, 
allowing researchers and developers to gain insights into the model's internal processing and behavior patterns. 
As an open-source tool, it serves as a foundation for advancing interpretability research and enhancing control 
over large language model operations.

__Model Creator__: [Goodfire](https://huggingface.co/Goodfire), built to work with [Meta's Llama models](https://huggingface.co/meta-llama)

By using __Goodfire/Llama-3.1-8B-Instruct-SAE-l19__ you agree to the [LLAMA 3.1 COMMUNITY LICENSE AGREEMENT](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct/blob/main/LICENSE)


## Intended Use

By open-sourcing SAEs for leading open models, especially large-scale 
models like Llama 3.1 8B, we aim to accelerate progress in interpretability research. 

Our initial work with these SAEs has revealed promising applications in model steering, 
enhancing jailbreaking safeguards, and interpretable classification methods. 
We look forward to seeing how the research community builds upon these 
foundations and uncovers new applications.

#### Feature labels

To explore the feature labels check out the [Goodfire Ember SDK](https://www.goodfire.ai/blog/announcing-goodfire-ember/), 
the first hosted mechanistic interpretability API. 
The SDK provides an intuitive interface for interacting with these 
features, allowing you to investigate how Llama processes information 
and even steer its behavior. You can explore the SDK documentation at [docs.goodfire.ai](https://docs.goodfire.ai).

## How to use

View the notebook guide below to get started.

<a href="https://colab.research.google.com/drive/1IBMQtJqy8JiRk1Q48jDEgTISmtxhlCRL" target="_blank">
  <img
    src="https://colab.research.google.com/assets/colab-badge.svg"
    alt="Open in Colab"
    width="200px"
    style={{ pointerEvents: "none" }}
  />
</a>

## Training

We trained our SAE on activations harvested from Llama-3.1-8B-Instruct on the [LMSYS-Chat-1M dataset](https://arxiv.org/pdf/2309.11998).

## Responsibility & Safety

Safety is at the core of everything we do at Goodfire. As a public benefit 
corporation, we’re dedicated to understanding AI models to enable safer, more reliable 
generative AI. You can read more about our comprehensive approach to 
safety and responsible development in our detailed [safety overview](https://www.goodfire.ai/blog/our-approach-to-safety/).

Toxic features were removed prior to the release of this SAE. If you are a safety researcher that 
would like access to the features we’ve removed, you can reach out at <a href="mailto:[email protected]">[email protected]</a> for access.